goal-driven-execution

Overview

Andrej Karpathy's key insight: "LLMs are exceptionally good at looping until they meet specific goals. Don't tell it what to do — give it success criteria and watch it go."

This skill converts vague imperative instructions ("make the login work") into declarative goals with concrete, testable success criteria. Agents with clear goals self-correct autonomously. Agents with vague goals produce vague results and require constant intervention.

When to Use

• Before starting any multi-step task
• When a task has been described imperatively ("do X, then Y, then Z")
• When you're unsure how you'll know when you're "done"
• For long-running or complex implementations

Process

Step 1: Extract the Underlying Goal

1. Read the full request.
2. Ask: What is the user trying to achieve, not just what they asked for?
3. Write the goal as: "The task is complete when [observable, verifiable outcome]."

Example transformation:

• ❌ Imperative: "Add error handling to the API."
• ✅ Goal: "The task is complete when: all API endpoints return structured error responses for 4xx/5xx cases, error responses include a code, message, and requestId, and the existing tests pass."

Verify: The goal statement is observable and testable by a third party.

Step 2: Define Success Criteria

1. List 3–7 specific, binary success criteria:

   Success when:
   - [ ] All existing tests pass
   - [ ] New behavior X is demonstrated by test Y
   - [ ] No regressions in file Z
   - [ ] Manual check: [describe what to look for]
   

2. Each criterion must be falsifiable — you can clearly state when it passes or fails.

Verify: Every criterion can be checked without the original author.

Step 3: Define the Execution Plan

1. Break the goal into ordered steps, each with its own verify check:

   1. [Step] → verify: [command or check]
   2. [Step] → verify: [command or check]
   3. [Step] → verify: [command or check]
   

2. Identify the first failure mode — what's most likely to go wrong? Plan for it.